Remote Emission Sensing (RES) enables non-intrusive, large-scale measurement of real-world vehicle exhaust emissions. However, linking RES measurements to vehicle technical data typically relies on license plate recognition, which raises data protection and data governance concerns. This study investigates whether vehicle groups defined by fuel type and emission standard can be inferred directly from exhaust emission signals, without relying on vehicle-identifying information, and how classification performance changes when repeated measurements of individual vehicles are available.

Vehicles were assigned to two groups based on the Real Driving Emissions (RDE) regulatory transition and observed emission behavior: pre-RDE diesel vehicles (Euro 5 and Euro 6a/b) and petrol vehicles together with post-RDE diesel vehicles (Euro 6c–Euro 6d-TEMP/6d). Four supervised machine learning classifiers were evaluated: Random Forest, eXtreme Gradient Boosting (XGBoost), Multi-Layer Perceptron (MLP), and Multi-Instance Learning (MIL). The classification was based on remote-sensing measurements of NO, NO₂, and CO. In the MIL framework, all measurements belonging to an individual vehicle were treated as a single bag, enabling direct vehicle-level classification from multiple observations. For the other models, measurement-level predictions were aggregated at vehicle level by averaging predicted probabilities, allowing scenarios with one to five measurements per vehicle to be evaluated.

The dataset comprised 438,225 measurements from 291,445 unique passenger cars measured in RES campaigns across Europe, covering diesel and petrol vehicles with emission standards ranging from Euro 5 to Euro 6d. XGBoost achieved the highest balanced accuracy among the evaluated classifiers, with performance of approximately 87% for single-measurement classification. Vehicle-level balanced accuracy increased to approximately 90% with two measurements and reached approximately 94% with five measurements. The largest improvement occurred between the first and second measurement, while additional measurements yielded progressively smaller gains.

The results demonstrate that vehicle groups can be inferred from exhaust emission signals alone, providing a potential privacy-preserving alternative to license-plate-based vehicle classification in RES applications. The approach was developed for passenger cars and has not yet been validated for other vehicle categories such as trucks or light-duty vehicles. Future work will investigate the transferability of the approach across vehicle fleets and measurement conditions and its integration into high-emitter detection frameworks.